Find file
Fetching contributors…
Cannot retrieve contributors at this time
499 lines (377 sloc) 17.2 KB

Tharsis: Threaded Entity-Component Framework in D

NOTE: This is a design document; it does not describe a finished product and is subject to change

Tharsis is an open-source Entity Component System framework written in the D programming language with support for automatic threading of game/simulation code.


Most past (and many present) game engines use an inheritance-based approach to design game entities. This approach is intuitive but gets unwieldy as the entities get more complex. This led to the adoption of entity systems based on aggregation instead of inheritance, many of which are referred to as entity-component systems. This has been discussed to death already, so this article focuses on the specifics of the Tharsis framework.

Many early component systems use entities composed of "intelligent" components with their own logic logic usually implementing some kind of an interface. This approach has various issues, including the dependency of a components' logic on another component.

Tharsis entities consist of "dumb" components that are pure data with no logic whatsoever, similar to the approach proposed by Adam Martin.

With this approach, all logic is separated into "systems" (or processes, as they're called in Tharsis). A process is usually (ideally) a stateless object with a method that is called for all entities containing components specified in its signature. This decouples data and code and avoids component dependencies; code depending on particular components will only execute if those components are present.

Another advantage of dumb components is performance; components can be plain-old-data types with no virtual functions and can often by tightly packed in memory.

Tharsis builds on this approach, adding a distinction between past and future state which removes dependencies of processes on the order in which they are being run and enables performance optimizations such as execution of processes on different threads.


Separation of data and logic

Components in Tharsis are plain-old-data with no logic. Logic is fully contained in processes, often called systems in other entity component systems.

Data-defined entities

Tharsis automatically generates code to load components from user-specifiable Sources (see below in the Concepts section). YAML support is built-in; support for XML, JSON or other formats can be added by defining a struct with a very simple interface.

Past/future distinction

Unlike in other frameworks, components and entities in Tharsis exist in two versions; read-only past and writable future. This removes the dependencies of processes on the order in which they run; all processes read the unchanged state from previous frame.

During a frame, processes read past components to write future components (which are created on-the-fly). When all processes have finished processing, the frame ends, future becomes the past, and memory formerly allocated for past is reused for future state created during the next frame.

Since future components are created as the process runs, the process can add or remove (i.e. not add) components while the data is stored linearly, avoiding dead gaps and subsequent cleanup. If an entity is removed, its components are simply not added In the future.


All processes in Tharis read immutable data from the past; one process can't affect what another reads during a frame. Also, only one process can write future components of one type. While this is somewhat limiting, it allows the processes to be moved into separate threads automatically, adapting to the number of cores of the CPU.

Generated code for safety and performance

Components in Tharsis are always structs and processes don't need to be polymorphic; Tharsis analyzes signature of the process() method of a process and generates code that will call it directly without indirections. This enables inlining and other compiler optimizations.

The process() method is only called if an entity has all past components from its signature. There is no need to cast or to determine if a component is present. Component iteration is not implemented by convention; it is built into the framework, allowing further optimizations such as threading.

Cache-friendly memory usage

Tharsis allocates memory through user-specified allocation functions (default being malloc/free). Memory for components and entities is preallocated based on compile-time hints (e.g. if a component type is common, one can specify that at least 16k components of that type should be preallocated) as well as run-time controls (e.g. increase the size of preallocated memory before loading a large map).

If Tharsis runs out of memory and is forced to reallocate in the middle a frame, it prints a warning, allowing the programmer to adjust size of preallocated memory.

Components and entities are tightly packed in simple arrays processed linearly to avoid cache misses.


This section describes the main types and concepts in Tharsis API. Note that the implementation may involve more types.

Component (concept)

  • Simple struct types defined by the programmer
    • Registered and verified by ComponentTypeManager
    • No elaborate destructor, copy constructor or postblit
    • Data behind any indirections is not owned by the component
      • E.g. resource handles (initialization not certain yet)
    • Can be safely compied by memcpy
  • Compile-time information:
    • Unique type ID (set manually by the user, validated at runtime)
    • Optional preallocation hints
  • Type name must have a "Component" suffix (may be relaxed if inconvenient)
  • Builtin components (avoided if possible)
    • LifeComponent - determines if an entity is dead and should be removed


struct StaticComponent
    enum ushort ComponentTypeID = maxBuiltinComponentTypes + 1;

    enum minPrealloc = 16384;

    enum minPreallocPerEntity = 1.0;

    vec3 position;

    vec3 rotation;

MultiComponent (concept)

  • A Component allowing more than one instance per entity - Differentiated by a compile-time flag or naming convention (not certain yet)
  • Passed to processes as a slice
  • Extra compile-time information: - Maximum components of this type per entity


struct ColliderMultiComponent
    enum ushort ComponentTypeID = maxBuiltinComponentTypes + 2;

    enum isMultiComponent = true;

    enum maxComponentsPerEntity = 1024;

    EntityID colliderEntityID;

Process (concept)

  • Class types with a process() method (without deriving a common parent)
  • Signature of a process() method:
    • Optional reference to immutable past entity
    • References to immutable past components
      • For MultiComponents, these are slices of immutable components
    • Optional output reference to mutable future componen
      • May be an output reference to mutable pointer; the pointer can be (and, due to the output reference, is by default) set to null to avoid writing the component to future state.
      • For MultiComponent, this is an output reference to a slice of mutable components with length equal to maxComponentsPerEntity of that MultiComponent type. The slice is shortened to specify the the part written to future state.
  • Multiple process() methods can be used to match different past component patterns (at some performance cost). If component patterns of two process() methods are ambiguous, a compile-time error informs the programmer about a need to define another process() function to handle a union of these patterns. More specific process() methods take precedence over more general ones.
  • A process may contain compile-time information related to scheduling such as the number of frames it can skip, or to allow optimizations in code generated to call its process() method/s. This information will be defined as the scheduling code is implemented.


class MovementProcess
    /// A more specific process() method, taking precedence if there is
    /// both a StaticComponent and a DynamicComponent.
    void process(ref immutable(StaticComponent) pastStatic,
                 ref immutable(DynamicComponent) pastDynamic,
                 out StaticComponent futureStatic)
        // How to determine time is not yet determined,
        // this is just an example.
        futureStatic.position = pastStatic.position +
                                timeStep * pastDynamic.velocity;

    /// A more general case, ensuring the the component is preserved
    /// in future state (this particular case may be simplified/optimized
    /// further with a compile-time flag if overhead is measurable).
    void process(ref immutable(StaticComponent) pastStatic,
                 out StaticComponent futureStatic)
        futureStatic = pastStatic;

Entity (struct)

  • Stores a unique ID (EntityID)
  • By itself, it has no data about the components it contains
    • Component counts are stored by EntityManager in arrays parallel to the entity array; a Process writes to the component counts array for its future component type, removing the need to lock entities.

EntityPrototype (struct)

  • Stores components needed to create an entity
  • Can be loaded from a file once, used to create many entities
  • Does not manage its memory (memory can be provided by e.g. a ResourceManager)

Source (concept)

  • Used by generated code to load Component properties
  • User-overridable (at compile-time) type to load components from (e.g. into EntityPrototypes)
  • Usually should be a wrapper around a serializing format such as XML, JSON, YAML. YAML support is implemented by builtin YAMLSource
  • Must be able to represent mappings, sequences, values

Resource (concept)

  • Struct types designed to be created once, used without mutation
    • May contain large chunks of data loaded from files, such as images, sounds, 3D models, EntityPrototypes, ...
    • Created once, marked immutable, reused many times from many threads
    • Must define two struct types: Descriptor and Handle
      • Descriptor contains all information needed to create the resource (e.g. a filename)
      • Handle is an ID used to access a Resource through its ResourceManager


struct EntityPrototypeResource
    struct Handle
        /// Accessible only by a ResourceManager in the same package
        uint resourceID_ = uint.max;

    struct Descriptor
        string fileName;

    this(const ref Descriptor descriptor) @safe pure nothrow
        this.descriptor = descriptor;

    // As the resource is accessed through immutable references, its data
    // members can be public. The resource manager can directly change them
    // until marked immutable.

    EntityPrototype prototype;

    Descriptor descriptor;

    ResourceState state = ResourceState.New;

ResourceManager (class hierarchy)

  • AbstractResourceManager - common base
    • Allows EntityManager to "run" the ResourceManager
  • ResourceManager!Resource - base of resource-specific managers
    • API to use Resources with
    • Resource creation, loading, access, state queries
  • Implementation classes
    • Implement ResourceManager!Resource API
    • Store resources in (logically) immutable storage

EntityManager (class)

  • The main "master" object
  • Entity creation (done between frames)
  • Entity/component storage
  • Resource manager registering
  • Process registering/validation
  • Executing processes
  • Configured at compile-time by a policy with limits, hints

ComponentTypeManager (class)

  • Component type registering/validation
  • Component type info generation
  • Configured at compile-time by a policy with limits, hints

Execution of a frame

This is a high-level overview of what EntityManager needs to do between frames. Note that this is very likely to change as any unexpected issues appear.

  • Wait until all processes running in threads stop (don't stop the threads)
  • Run resource managers (allowing them to load resources, possibly in background threads)
  • Switch past/future component buffers
  • Reallocate memory if running out (this should not happen often)
  • Create entities added during the last frame
  • Switch past/future entity buffers
  • Copy the past (former future) entities to future, forgetting dead entities
  • ? Handle messages ?
  • Assign processes to threads
  • Run processes in threads

Execution of a process

This is a high-level overview of what the generated code executing a Process needs to do every frame. As above, this is very likely to change.

  • For each past entity
    • If the entity is not dead
      • If the entity has all needed components
        • Ensure we have enough memory (running out of memory here should be very rare, and this check could even be removed based on preallocation hints).
        • Call (one of) the process() method(s)
      • Update the component count on the future component
      • ? Optionally update a spatial manager based on this component ?
    • Iterate over components relevant to the process

Uncertain features / to be designed

  • Message-passing between systems (always delayed by 1 frame). - May not be needed at all
  • Spatial management (needed for e.g. collision detection) - Should be user-defined


  • Main goals
    • A past/future entity system
      • Separation into past and future state
      • Future state is generated on-the-fly
      • Resources loaded async between frames (but not in separate threads)
      • Entities loaded from files using compile-time generated information, including resource initialization.
      • Performance measurements
    • Threading
      • Processes executed on separate threads
      • Solving uncertain/unforseen issues.
        • Messaging ?
        • Spatial management ?
      • Performance measurements
    • Scheduling
      • Processes assigned to threads according to user hints and/or (very basic) real-time profiling
      • Performance measurements
    • Optimizations
  • Optional goals
    • Threaded resource loading - Resources loaded async in background threads
    • Special-case optimizations in generated code

Non-goals (aka beyond-1.0)

  • Parallelization on the level of entities as opposed to (or in addition to) processes
    • Past-future distinction is the main requirement enabling this approach
    • Possible, but would complicate the design (memory organization)
    • Not very useful in most cases on <10 core machines
  • Parallelization on the GPU
    • Will require massive changes
    • May be revisited once unified address space is common
    • May only be useful with entity-level parallelization
  • Scripting language support

Other component-entity system frameworks